pakl

What was the architecture of your predictive model? Was it designed to learn the underlying physical dynamics from the tons of sensor data?

May I ask how you reached this insight? What field do you work in?

> objc_msgSend is written in assembly. There are two reasons for this: one is that it's not possible to write a function which preserves unknown arguments and jumps to an arbitrary function pointer in C.

Wow... this is a bit off topic but can anyone expand on this side note and explain why?

(Every Objective-C implementation requires assembly code?)

In the interview, Tim Cook says nothing about working on cars. He recognizes that a new core technology is needed for autonomous systems, but doesn't say anything to imply they're working on a car.

Along with Cook recently saying humans should be involved in decision-making, there are some very interesting AI-related statements coming out of Apple these days.

If this interests you, you should also check out carbon co2, a lightweight object-oriented addition to C (https://github.com/peterpaul/co2/).

Code example: https://github.com/peterpaul/co2/blob/master/examples/my-obj...

If you actually try current vision systems out on real raw video data as opposed to clean datasets of "good" photos pre-selected by humans, you'll see that they are terribly far from human performance.

Same goes for translation systems.

Most current systems (by their very design) lack dynamical representation capability necessary for modeling interactions in/of the world. I hypothesize this is important for AI that actually gets what it's dealing with.

It's also stunning how poorly it performs in a real world (i.e., non-human-prefiltered photo) scenario.

Like autonomous cars or security cameras or robotics.

Change to non-canonical lighting and it fails on classification tasks.

Add non-standard background noise, and it also fails on hearing tasks.

It's also stunningly poor in lots of cases. Largely because the system has no idea of the reality it's talking about.

I'm certainly not saying that deep learning is fundamentally flawed. It's a great method, very powerful. (Excellent algorithm.)

I'm saying it's not reasonable to expect good generalization in deep convnets that learn mappings from static images to human labels. (Wrong problem.)

Training on adversarial examples doesn't solve the fundamental problem, it merely tries to plug the holes. But in such high dimensional spaces there are many many holes to be plugged. :)

Agreed the failure mode may seem esoteric, but note that OpenAI is making a big deal about them.

A non-esoteric way to demonstrate the lack of generalization is to feed a deep conv network real world images (from outside the dataset). Grab a camera and upload your own photo. Roboticists who try to use deep conv nets as real world vision systems see these failures all the time.

> I have this diffuse idea in my head that most possible images do not occur in the real world and that there are way more degrees of freedom into direction that just don't occur in the real world but this idea is just too diffuse so that I am currently unable to pin and write down.

Yes! You're on the right track! The number of degrees of freedom of images of pixels and textures is HUGE. There is not enough data to practically learn directly from those images. So the deep networks are starved for data -- even with the big datasets they are trained on. (It's only thanks to the way they are set up they do well when tested on very similar images, like sharp hi-res photos. But they fail to generalize to other kinds of images.)

So how can you reasonable reduce these degrees of freedom?

It turns out that the continuity of reality itself provides a powerful constraint that can reduce the degrees of freedom. See, when a ball rolls along, this physical event is not just a collection of textures to be memorized. It's an ordered sequence of textures that vary in a consistent and regular way because of many learnable physical constraints (like lighting).

So, it turns out you can reduce the dimensionality by making a particular kind of large recurrent neural net learn to predict the future in video. Our very preliminary testing shows it works shockingly well.

Adversarial examples are just one way to prove that deep learning (deep convolutional nets) fail at generalizable vision. It's not a security problem, it's a fundamental problem.

Instead, ask yourselves why these deep nets fail after being trained on huge datasets -- and why even more data doesn't seem to help.

The short answer is that mapping directly from static pixel images to human labels is the wrong problem to be solving.

Edit: fixed autocorrect typo

Alternative TL;DR: The experimental methods you use can color the results you get.

E.g., lesioning different parts of a microprocessor may reveal "clusters of function" much as such studies do in the brain.

Does anyone know the source for his Einstein quote on page 9: "You should be careful to distinguish what is true from what is real."?

I believe Objective C's @property declarations brings some level of what you describe to C/C++. They generate the boilerplate for you.

E.g., @property (readonly) NSString* myName;

People do write those, and even propose viable alternatives, but those articles don't get upvoted as much ;)

If the shadow is predictable then the unsupervised (self supervised) part of our Predictive model will not have too much trouble learning to deal with it. But you're right, if it did have trouble, there would be good reason to for the system to continue exploring that type of event.

One way to do this is to link PVM'a prediction error output to an "instinct" that directs it towards lower confidence events that you mentioned. This could be just orienting the camera towards those events, or in a robot producing actions that led to those events again.

Note this is not necessarily reinforcement learning but it relates to some ideas there, like the idea of novelty being rewarding.

Yes, that's right! The way you are using supervised learning here will force the neural networks to map from textures directly to human labels. A purely feedforward network, no matter how deep, can only rote memorize the effects of the world on the images (viewing angle, lighting, etc) and will not generalize.

Another shortcoming of feedforward nets is they cannot change how they interpret local features based on integrated global aspects of a scene, like ambient lighting or backlighting.

As a result the network will fail to classify on new real world images.

If instead you use recurrence to learn features that take the dynamical and global effects into account, you'll have a better chance of success. One example of how we did his is here [1].

[1] http://blog.piekniewski.info/2016/11/04/predictive-vision-in...

Oh-- my comment wasn't about the simplicity of the first part. (In fact, this is a great tutorial, thanks for posting.)

My comment is about the approach of using supervised learning to map directly from images to category labels.

If he ever tries to train deeper models in this manner and test them on real video frames from a car, he will be in for some unpleasant surprises.

Learning to map discrete snapshots of objects to labels won't yield a system that can deal with e.g., the reality of lighting conditions that a car will experience.

Working on self-driving cars, much like working on advanced robots, requires tight integration between the software, sensors, and motors. These aspects cannot really be developed independently, especially if you want to end up with an optimized system that could be a product.

Safety is a part of this, too.

Google has ~10 years of experience in self-driving cars. They also have acquired many robotics companies. Surely, they know this.

[edit: reorder]

This is a good point, I didn't know Alan Kay had said this.

Logic can sometimes be used to describe (to model) some minute aspect of thinking. But it is nearly always a massive simplification, discarding much of what happened during the thought process.

A couple of similar examples that might help:

- Thinking is not arithmetic (but arithmetic can be used to describe the result of some thoughts about stuff.)

- The solar system is not computing anything (but we can use the patterns as a though it were a program executing a timer/counter).

A related point is that thinking is not "executing an algorithm on a Turing machine".

Lecun has identified a real problem for AI -- the need to understand the real world, the link between intelligence and prediction over time. But the tools he is using are not the right ones.

Deep conv nets were not designed with prediction over time in mind. Here's one reason: Deep convolutional nets do not handle dynamical information from their lowest layers. By design, conv layers and pooling layers immediately begin discarding spatial information that could be used for building up predictions.

In contrast it is possible to start with recurrent/feedback networks at the very first layers of the network. These initial layers can begin building up predictions at the pixel, color, and lighting level (example: our recent preprint[1]).

My colleague, who as some of you know enjoys blogging, wrote a more thorough post in response to LeCun's recent CMU lecture on the same topic as these slides[2].

[1] https://arxiv.org/abs/1607.06854

[2] Blog: "A few comments on the Yann LeCun lecture at CMU, 11.2016" http://blog.piekniewski.info/2016/11/21/yann-lecun-cmu-11-20...

I work at LeEco US out of San Diego, and my colleagues work at other ML/AI companies also in San Diego. We originally met and collaborated at Brain Corporation.

In a sense, I would say yes there are learning laws, but it's still early in codifying them.

Along one axis, you could compare: supervised, semi-supervised, self-supervised and unsupervised learning. Along another axis, consider that there are versions of each method that take into account temporal/dynamic data, versus others that require randomly shuffled static data.

In the current problems of visual perception, I think the field would benefit greatly a shift to focus on multiscale interaction/dynamics rather than on (static) statistics as it is currently (for more on this, see my colleague's blog: [1]).

[1] Statistics and dynamics. (http://blog.piekniewski.info/2016/11/01/statistics-and-dynam...)

If the goal is to have AI that is aware of the real world (even a chatbot), then using game state is a crutch that doesn't help us solve the real problem.

That's a great (and hard) problem!

More generally, imagine AI that could learn the physics of the world. For example, if the ball is rolling away, the AI should be able to predict that the ball will look smaller on the next frame.

Going further, if the ball is about to roll under a shadow, the AI should predict that the ball will become a darker shade of green.

(After several years working in a robotics research company, these kinds of capabilities are exactly what we determined would be necessary for robot AI.)

I take the point to be that there aren't "deeper" fundamental principles at play in these models. Tremendous progress has comes from simply tweaking of the numbers of layers, or how the feed forward to each other (skipping layers, etc), or by throwing more computer power or data at the same basic algorithm.

Where might we look for deeper principles? One idea is to consider what brains do and how they might be doing it. (I'm not saying we need to go down the rabbit hole of biological detail -- on the contrary I'm suggesting we look at known or even hypothesized principles of brain operation and import them into AI.)

Two ideas we have used in our work: prediction (over time), recurrent feedback (most brain regions have more feedback than feedforward inputs)

What you mention in point #1 isn't the main issue with games, but with the "reactive" approach in AI. Reactive AI systems are necessarily going to be more limited than say, predictive AI systems (a new emerging class of AI models). [1]

Agreed, though, that the focus on games is a long-standing source of problems holding back AI. My colleague has referred to this as "ludic AI" in a recent blog post where he expands on your statement about the real challenge of intelligence. [2]

[1] Reactive vs Predictive AI (http://blog.piekniewski.info/2016/11/03/reactive-vs-predicti...)

[2] Learning Physics (http://blog.piekniewski.info/2016/11/30/learning-physics-is-...)