So long as you don’t dispense medication (requires measurement) or look at clinical trials for very sick patients (requires understanding statistics) or do research in applying machines to data interpretation (helps to know calculus) or want to talk to all your other physician colleagues (who did well in math), then you should be OK.

Answer 2: There is an MIT AI Memo or TR from about 1977 that discusses machine generated disney-style story generation. The “princess” usually appears on stage with grand sweeping motions and the villain also has his own gestural identity. If the robot can recognize gestures, it could act to fill in missing roles. E.g. If the robot recognize the gesture of the person in the room as being a villain, it could become a princess and sing “some day my prince will come..” and move back and forth in sweeping gestures. If, on the other hand, it recognizes a princess coming on stage, it could adopt the role of villain. If you create something that is solely a princess, it could wear a dress and spin faster or slower based on the speed. Be sure to use a high pitched cartoon voice. The effect will be better.

Simpler to implement, have an LCD “face” that brings up different disney princesses depending on the aria sung by the villian. E.g. if it recognizes Gaston, the face of “Beauty” pops up. I say this assuming that spectral analysis show easy to identify differences between the different villains. Step zero in this my be to have the villain announce themselves. “Beauty, it is I, Gaston!”. Worse case, you record the audio from the film and use that as the trigger. The nice thing about this is that there are so many villain/heroine pairs out there that you can try all of them and see which two your recognition algorithm finds farthest apart. So.

1. LCD face + Dress + Speaker (High quality in the upper frequency range. Needs to render soprano perfectly). Dress can conceal wheels that make it spin. as it moves.
2. SImple Audio villain (hero)/ recognizer (from film’s audio tape)
3. Swap out wheels for legs
4. Improve recognizer to recognize name instead of villains audio spectral signature.
5. Have the robot recognize its ‘cue’ and sing only after it recognizes it.

You have now created a mechanical actress that can play any disney role, will not forget its lines or step on the lines of other actors.

=======================================================

Answer1: If the drone is airborne, you can probably accomplish most of this by following the script given in aero_astro 16.30-16.31.

16.30 is s semester course, so make sure you have enough in place that you don’t have to recreate everything from scratch.

Is the input verbal? If so, does Siri have an API? If so, put a fair amount of thought in error handling. E.g. if your professor tells it to “go to hell”, you could say, “Thanks, but I’m already enrolled in your course”.

Perhaps the easiest thing to do, given that the semester is started, is to break your project into demonstrable steps:

1. use a laser pointer to specify the destination, and have the drone lock on that.
2. Once that milestone is reached, you can put the dot on a moving object (a car, a human head) and have the drone follow it.
3. Once that milestone is reached, you can fly over the object and drop a pellet on it.

Bill Paseman's answer to Life originated from chemicals, Abiogenesis, and evolution did the rest of the work. All of our thoughts and emotions are simply electrical charges and chemical reactions in the brain. Are we just bags of chemicals, moving around, and nothing more?

Though not calculus specific, I enjoyed The Mathematical Experience: Phillip J. Davis, Reuben Hersh: 0046442929684: Amazon.com: Books

Judge Schonfeld (Real-Time Evidence Based Medicine) announced in November 2015 that his group had discovered that a cancer drug could also be used to treat Ebola.

This leads to the world of Drug discovery by Drug repositioning - Wikipedia. The basic idea is that there are lots of drugs out there that have passed phase 3 trials (I.e they don’t kill you outright, dosing is known and they have some non-fatal effect on the body) and all these drugs have side effects (The most famous example is Sildenafil - Wikipedia, which originally was used to treat pulmonary arterial hypertension, but was then discovered to treat erectile disfunction and was rebranded as viagra.)

So you put all these drugs in a database and run algorithms to see if they can be used to treat conditions other than the primary target. This is an area of active interest since 25% of most drug company revenues come from repurposed products (Therapeutic Drug Repurposing, Repositioning and Rescue Part III market exclusivity using Intellectual Property and regulatory pathways.)

There are lots of ways to do this at many different levels (e.g. [1502.02072] Massively Multitask Networks for Drug Discovery). However (I suspect) most of this work is being done internal to drug companies since they are most interested in leveraging their own drug portfolio, not curing a disease using someone else’s drug.

Neural networks are black boxes, They are an almost total surrender of mechanistic reasoning to statistics. The explanations then of why they work (or don’t) are based on statistics, not aristotelian logic. From the video below: “the sample set is too small.” “The bias is too big.” “The variance is too big.” Not “The system does not understand green chihuahuas in the rain”.

As such, I suspect your question is non sequitur, the designer -never- understood the reason why the model worked well (or not) in the first place. At least, not in non-statistical terms.

The network develops its own way of answering the question (Note in particular Dr. Ng’s “Phonemes are a fantasy of Linguists” comment. ) A wonderful example is from Detecting Tanks, where Ed Fredkin saw a presentation describing a network trained to detect tanks, and remarked that the network actually was distinguishing cloudy and non=cloudy days.

This video is listed here: Bill Paseman's answer to What are the most important deep learning algorithms? In which order should I learn them?

When it starts paying taxes

rhiever/tpot

Check out protégé

Bill Paseman's answer to What are the most important deep learning algorithms? In which order should I learn them?

I think it is a reasonable question to ask.

There are lots of aspects. E.g. I went on a Boeing tour several months ago, and they showed their assembly line of planes. The newer ones are smaller, made of composite materials (vs. aluminum) and have computer controlled devices everywhere (e.g. the passenger window glass).

Hobby Drones are steadily improving. Go to Fry’s and look at the latest model features and compare to some of the older models. This will give you an idea of feature trends.

Tesla develops drone components for its cars (electric motors, navigation and batteries). Look at projections for improvement from the analyst reports and assume the components can be repurposed for drones or light aircraft. This will give you an idea of component trends.

I was a grad student in Steve Ward's group at MIT from 1978-1980.  My Masters thesis explored the application of AI to music composition.  Although I framed my work in terms of generative grammars, a colleague, David Levitt noted that jazz improvisation could be structured using Waltz-like Constraint networks.  I spent a few weeks adapting this to species counterpoint.  Fast Forward to 1992.  I had a contract to create a "configurator" for WANS, and later PCs.   I noted that I could adapt the same constraint code to configure these products (as well as insurance policies, financial plans, fork lifts, etc.).  I formed a company called Calico which commercialized this.  It went public in 1999.  I then retired.

Two Notes:

1) I actually have degrees in ChE (Petroleum Engineering) EE (Signal Processing) and EECS (AI).  And after all of that, applying AI to music composition is what made my fortune.

2) Probably having a broad background helped me be a better entrepreneur than knowing AI.  My attitude was that I could pretty much do any type of engineering.  Engineering is the commodity.  What is in short supply is good, expensive, solvable customer problems.  A broad background allowed me to ask the important question: "What is your most expensive engineering problem and how much are you willing to pay to solve it?"

Its been done. Long ago.

When PARRY Met ELIZA: A Ridiculous Chatbot Conversation From 1972

Joe Weizenbaum wrote the original ELIZA. I actually had his office after he left.

He once read a gushing review of “Parry” (above) that said that PARRY provided a wonderful opportunity to study the mechanisms of paranoid schizophrenia .

Dr. Weizenbaum wrote a biting response. He said that he had just created a program to simulate a far worse malady: autism. It consisted of one line:

While (True); input X

He mentioned that it was so good that it would function even if the computer wasn’t turned on.

This anecdote does a good job of pointing out the difference between simulating the I/O characteristics of something and creating a higher fidelity model.

Critical to note though that even creating a high fidelity model is not the same a creating a ‘being’.

Well, how deep a knowledge?

Let me tell you a problem I have, and if you can solve it, let’s discuss a price.

Home - ClinicalTrials.gov has a database of semistructured data. I need to narrow down the descriptions of particular conditions to a small selectable subset. Currently, I have to hand code a recognizer for every single goddamned disease set. And my work is not that good. E.g. if the text says “you can participate in the trial unless you have condition ‘X’”, my code happily says that people with condition ‘X’ are eligible for the trial.

So. If you can do stuff like that, then I think you have the basis for a good consulting practice.

Re PC, Yes, if you believe the Church–Turing thesis - Wikipedia (which has been around since the 1930’s) and speed is not an issue.

To Emmett’s point, a -natural- model of human intelligence has not yet been puzzled out. Speculations here Bill Paseman's answer to Life originated from chemicals, Abiogenesis, and evolution did the rest of the work. All of our thoughts and emotions are simply electrical charges and chemical reactions in the brain. Are we just bags of chemicals, moving around, and nothing more?

For each initial (multi-million) investment I got for my two companies, the funders (Kleiner perkins and Mayfield) sent a doctor to my office to do a physical which included a blood and urine test.

See What technology should we expect cities to realistically adopt within the next 5-10 years (2021–2026)?

An elaboration of Daniel's point.  An idea does not have to be in steps.  But a BIG idea is most likely implemented in steps.  There is an old quote. "Don't mistake clarity of vision for shortness of distance".  You may be able to clearly see a mountain on the horizon and decide that you want to walk to it; but if it is hundreds of miles away, miles high and there are several deserts between you and it, you most likely have to have a plan beyond just "walking" to reach it. 

Since things like ideas, plans and steps have a tendency to be confused, and this confusion results in lost time and effort (especially among groups), Planning methodologies often formally distinguish them.  For example, from GOSPAPlanning.com UK, US, Europe

• A goal is something we aim to achieve, an end result.  For example quality service, market leadership, number one sales person or employer of choice.  Our goals are our motivation.
• An objective is a deliverable, it is measurable and within a timeframe. For example $23m by year end, 10 sales meetings per month or 90% staff utilisation per month.  Objectives support goals.
• A strategy is how we are going to achieve our goals, these are what we should do.  For example implement a quality review process, add extra product functionality or increase product marketing in a specific vertical sector.  Strategies support Objectives.
• Plans are specific elements of the strategy.  They are what we are going to do in the short term (next 30 days) to move the strategy on.  For example select a marketing agency by the end of October or complete the product specification by October.  Plans support Strategies.
• Actions are day to day activities and are specific details of a plan.  For example ring Bill at Alpha Marketing Services.  Actions support Plans.

No.

It is a conspiracy of linen weavers.

If you have a degree of some kind, you probably had to focus to complete it. Whatever methods you used then, try applying them now.

If not, create a plan for each project and set small milestones. Try to arrange the milestones so that the key feature you want to demo is available after a short period of time. Continually demo to an external party and see if you still feel good about the project after finishing the demo milestone. You may find these three examples helpful.

Windows: In the early 80’s at Daisy systems, the president asked me to create a windows system. The OS group had several people working for a year and hadn’t completed anything. Much of the work in windowing system was handling updates to overlapping windows. So the first thing I did was force any window that was updating to the front so I did not have to deal with the problem. I also had my window system only work inside of one application (a discrete event simulator) so I didn’t have to deal with multiple processes. I leveraged existing packages (a bit level “screen region” mapper and a symbol table package). I had the first version working in 3 days. The president liked the crude demo so I spent more time on it. Adding borders, providing control key commands to grow/shrink/create/delete windows and associate signals with them, etc. That took 3 weeks. The next demo went well and I turned it over to another guy to add a test suite , fix bugs, etc. It took him 3 months. This window system-inside-a- simulator worked well enough that the sales team could sell it. This gave the os group time to finish their version.

Simulator: I had gone through this 3 day, 3 week, 3 month cycle before while managing the creation of the discrete event simulator. The simulator had three main parts: a console, a simulation engine and a compiler that turned the schematics into something that the engine could simulate. I first had to create a command console, so I reused the schematic editor console that I had spent several months on and simply switched out the schematic editor’s command parser. That took one day. The one-day demo simply consisted of entering a sequence commands that we would eventually have. The commands did nothing. I demoed it to the VP of Engineering and he was very happy since although it did nothing, it did it after just one day. I then hard coded a simple 4 nand-gate xor circuit for the engine to simulate and added debug commands so the engine developer could dump the network and queues. That took three days. (Another lesson here is that you will need a specialized debugger/structure dump for any complex product implementation. The earlier you build it, the more you can use it. So build it first and continually add to it.) Hard coding the xor circuit into the simulator decoupled console/engine development from the compiler development and let me demo the console and let the engine developer debug his engine before the network compiler was even coded. The next demo simulated a few steps of the 4 nand gates and did a hex dump of the state array to the screen after each step. No schematic compiler at that point. The first compiler took 3 weeks. Three months to work properly (During that time, the engine added additional behaviors beyond ‘nand gates’ and the console moved from a hex dump to displaying waveforms.)

Spreadsheet: I once had a consulting assignment where I came into a spreadsheet company which was in complete disarray and had had a series of VPs of engineering, all loathed by the staff. My mandate was to get things on track. They wanted me to be the next VP. I said “no”. I asked for a VP’s consulting rate, but asked to be assigned to the “in memory database”. Since the DB had no user level visibility, I politely asked each of the other groups: UI, reasoning engine, etc. to please provide me with an interface so that I could demo my db to the president. The first demo was simply to bring the product up, show a prompt and quit. (They hadn’t even gotten that far yet.) But this demoed a complete product integration. Next we showed a few cells that came from the database. And so on. My biggest contribution here was to identify the two key -emotional- issues: 1) The developers hated bosses, so I didn’t become one and 2) ultimately, they wanted to show off their work and demos provided an opportunity to do that -in the context of a whole product-.

Priority Algorithm: Now, here is the priority algorithm. Create these 3/3/3 schedules for all your projects. The ones you can’t do this with, Defer. Demo the 3 day milestone for each project and see which one you (or your customer) gets excited about. Again, the project’s one killer feature needs to be there after 3 days. If you can’t conceptualize, implement, demonstrate it; find a project you can. The implementation will have a LOT of stubs. resist the temptation to fill them out for the 3 day milestone. Again, the idea is to demo the one ‘killer’ feature.

Philosophical BS: Jim Barkesdale (CEO of Netscape) used to say “Never confuse ‘clearness of vision’ with ‘shortness of distance’”. When traversing that distance, recognize that the key issues here are emotional, not rational and that people (including yourself) can maintain faith and hope for only so long. So focus on short, clear milestones to justify faith and increase hope. Past the three month milestone, you still need to set proper milestones. You might find this helpful: How do I stay motivated and on track as a 'lone wolf entrepreneur'?

Just saw this quote from Napoleon Bonaparte: “A leader is a dealer in hope.” In my experience, this is true even if you are just leading yourself.

You could use them to analyze seismic logs, well logs, satellite photos. Any non-structured visual or auditory information.

You might try coming up with several demonstrable models and put a placard in front of each to show the resource flow and the math. Here are a few.

Amazon.com: SodaStream Jet Sparkling Water Maker Starter Kit, Black and Silver: Kitchen & Dining - How much CO2 can you actually pump in per unit volume? What happens when the container is breached? How long can you keep the container under pressure before “natural forces” cause it to leak away its contents? How much CO2 out there must be sequestered in order to change climate dynamics? How much CO2 can all the suitable aquifers are there actually accommodate? How much energy is required to do this? If the energy is created by burning carbon, how much coal/oil/natGas efficiency is lost by CO2 sequestration respectively?

Creatures that create shells vs. making concrete - How hard is it to sequester CO2 as calcium carbonate? Much of that comes from limestone, which comes from sea creatures. This form of sequestration is very stable, but takes longer than the other approaches. In fact, a lot of CO2 comes from decomposing limestone into cement using heat. (Cement is used in making concrete, which also generates CO2). Suppose you simply made less concrete, leaving existing limestone alone. From a chemical balance perspective, you could view that limestone as being “pre-sequestered”. So how much “pre-sequestration” would you get if you simply replaced concrete with another building material? E.g. asphalt roads (which also generates CO2).

Grass vs. trees - I read somewhere that prairie grass is more effective at capturing carbon than a tree is. However, both eventually die and, at some point, release CO2 back to the atmosphere. This is happening to the tundra in Alaska, which is adding to the CO2 load now that the snow cover is gone. At what rates can you sequester CO2 for each approach? Both need water. How much energy do you need to get them water, and how much carbon dioxide is generated by creating that energy?

In doing the balance equations, for non-carbon based energy generation approaches, how much energy is used to create a solar panel (be sure to factor in a panel’s usable life) or the hoover dam (made of concrete- see above) or a pound of U232 by the TVA? I once heard that if you looked at the energy balance eqns used to separate U232 using hydro-generated power by the TVA, U232 was not so much an energy source as a battery. I.e. it took 1 kw of hydro energy to generate enough U232 to generate 1kw of power.

You could potentially use it to track illegal bitcoing commerce. Eventually discovering who bought what for how much.

Bill told Hilarie to pay attention to Wisconsin. She and her advisors laughed him out of the room. They felt that he was concentrating on his voters, not hers. So here is an ai. Write it on a piece of paper and hand it to the next democratic presidential candidate. “Shut the hell up and actually listen to what bill Clinton has to say”. Perfect piece of code. No computer required.

I don’t know. My kids participated in several science fairs and I know that some projects were trendy. As I recall, the winner had an idea of how to use orange peels for water retention, yes? I can see how that would appeal to judges in drought stricken California.

However there are many examples of unappreciated ideas which worked out well (e.g. FedEx). If you believe in it, you might show it to others and see if it can be explained more clearly. Then extend it and resubmit it to another science fair. If it is really good, turn it into a company. Who knows. Maybe it will be acquired by Google.

Young Michael Dell (Dell Computers founder, author of Direct from Dell: Strategies that Revolutionized an Industry (Collins Business Essentials): Michael Dell, Catherine Fredman: 9780060845728: Amazon.com: Books) aggregated paper routes by opening new territories, recruiting subscribers, staffing the routes and taking a monthly cut when collecting fees.

He got in a conversation with a school teacher who saw him drive a BMW to school. Turns out Michael made more per year than the teacher did.

The teacher was upset, but “took him seriously”.

Net, do something entrepreneurial that makes you more money than them.

What are the most important deep learning algorithms? In which order should I learn them?

There are several organizations which help pay for surgery if you can’t do it yourself. Forums like SmartPatients discuss questions like these. Get your sister to reimburse her surgery through these channels and pay you back.

Keep meticulous track of the costs and payment sources and write a book about it, offering advice to others on how to get back on their feet.

If it is it a hobby, then you’d probably be funding it yourself or via crowd funding.

If you want to sell a product or service, I did this by starting a consulting practice to gather requirements, create a prototype and get initial funding (from the consulting fees). Once I got traction, I got VC funding. Teknekron was expert at this. Checkout the Harvey Wagner references here: P&A: Pitch Checklist

Build software better, together

OK, From the comment, the question is not about sharing, but about promotion.

Not my area of expertise, but traditionally, you’d

• get your own website
• Hosting, Domains, Website Services & Servers
• Domain Names | The World's Largest Domain Name Registrar - GoDaddy) ,
• instrument it (URL builders - Analytics Help),
• track it and
• promote it (Primer | Marketing Lessons from Google).

check uut Bill Paseman's answer to What are the most important deep learning algorithms? In which order should I learn them?

The best resource listed there for introductory machine learning is Andrew Ng’s coursera course.

1. Effectively represents whole sentences as vectors
2. Captures compositional meaning
3. Can be applied to images and sentences -jointly-.  This allows him to automatically generate explanations for images.

A neural network is usually initialized to random values before applying gradient descent algorithms.

Robert Taylor is correct; but given that…

The way I made the decision was to pick the job that trained me to do what I wanted to do in a few years. In my case, I passed up a bigger salary in exchange for exposure to how a startup worked. That approach worked well for me.

Here's one: Large-Scale Machine Learning for Drug Discovery

By building stuff I was interested in.

At the time (50 years ago) I lived in Houston, so I figured out how to make kites that I flew during hurricanes that could pull me very fast on my wagon. As I got older it was stuff like blue boxes, red boxes and black boxes (Blue box - Wikipedia). Then programs.

In software, here are examples of how i figured out to make stuff fast (Bill Paseman's answer to Every time I begin a personal project with an idea, I fail at continuing it. What can I do so I won't change ideas fast?)

Now I’m working on stuff using jet engines.

For what its worth, I recently heard a Cornell MechE professor advise students to work in a bike shop.

My first startup was created around a product I loved.  It failed.  My second was created around a product that customers would pay dearly for.  It succeeded.  If you can tie your technology to a problem that people will pay to solve, I suspect it will have a better chance of success.

Amazon.com: The Complete Handbook of Science Fair Projects (9780471460435): Julianne Blair Bochinski: Books