Is It Deja Vu All Over Again? (Part IX)
What’s It All About, Alfie? (Part IV)
Summary
In order to better understand current research in neural networks for AI, we explore the analogies between recent developments in this area, by comparing them with the comic and comically complex Rube Goldberg machines (RGM) of another era. Both are concerned with using machines to mimic human behavior. By comparison with the latter (RGM), we can shed some light on the former (neural networks in AI) and have some fun in the process. In the following, the term “neural network” (NN) refers to an artificial neural network realized on some sort of computing device being used for artificial intelligence (AI) purposes.
Background – Machine Learning and Deep Learning
Early in the development of AI, the basic neural network (see our First Dialogue) was to be the real work horse of artificial intelligence. It fell into disfavor for various reasons, on and off again over the years since 1956. Fairly recently – probably beginning around 2005, as other approaches (expert systems, etc) had temporarily taken their place in importance, neural networks assumed a new level of interest. This was due primarily to advances in computer power and access to very large data sets which could be used for training.
Today these neural networks, together with suitable mathematical algorithms for “training” them, form the most common basis for what is being called “machine learning” and “deep learning”. These latter objectives, now dominate much of the current work being pursued and reported in the AI community. Machine learning (using typically less than 3 hidden layers), and in particular, deep learning (more than 2 hidden layers), have gained great visibility in various social and commercial applications. These include surveillance (facial recognition, etc), and decision making (whether or not a candidate qualifies for a loan at a bank, etc), for example. Today, deep learning dominates the more successful efforts that we read or hear about.
Yann LeCun, the director of AI for Facebook, in an IEEE Spectrum interview, has compared a typical deep learning system to a box with possibly 500 million knobs (parameters) to be adjusted, or tuned, using readily available mathematical algorithms to accomplish this (in AI terminology, this is referred to as “training” the network, or simply “learning”); 1,000 light bulbs which will turn on or off as output signals from the result of “training” (or tuning) a network (or system); and 10 million training examples (which may be images, for example) to be used as example inputs for training the network to recognize or identify something of interest to us.
The inputs are used to train the network (or system) to achieve some goal related to the input training data. For the sake of discussion, we shall assume that the goal is to determine whether or not a particular object is present in an input image. After training the system, using perhaps millions of examples of images which do or do not contain the object of interest, the light bulbs (outputs) should light in a desired pattern when presented with any of the inputs used to train it, which do contain the object. Otherwise the desired pattern should not be seen. In the simplest case there might be only one output (one light bulb), and we hope to train the network to light up the bulb (or output the value 1) if the object is present. Otherwise the bulb should not light (output the value 0). The hope is that when the network is presented with a new but possibly different image containing the same object, as input, the network will yield the correct output (1 or 0), depending on whether the object is present or not.
The Problem and a Warning
The original basic neural networks (see our First Dialogue) which are used for machine learning (few hidden layers), and deep learning (more hidden layers), have not always worked out very well in practice, and in particular, have failed to produce anything even close to artificial general intelligence (AGI), also known as strong AI.
At the time of this writing, a trained neural network is generally only useful for the application it was trained for, and, in general, it can be so sensitive to small variations in the input(s), that even minor variations in an input can sometimes result in an incorrect output. In fact, it is possible that not all of the outputs will be correct even for data included in the training set. This problem severely limits the usefulness of the network to perform the task for which it was trained, but it also brings into question the credibility of any output which the network produces.
In some cases, this weakness cannot or should not be tolerated due to the serious nature of what the result might be used for. As examples, you probably don’t want such networks to be making decisions related to driving your car, or deciding your guilt or innocence in a court of law. Even if the network often gives a correct result, you can never know for sure if a particular result is going to be correct, or you may find out the hard way by suffering the consequences of an incorrect output.
Modifying Neural Networks to Improve Performance
As a result of the failure of basic neural networks to achieve a higher level of performance in machine and deep learning, practitioners of AI have begun experimenting with various ways to modify the original structures of the basic neural network by adding more and more elements. These are usually borrowed from respectable scientific fields such as signal processing, pattern recognition, and control theory, for example. The goal is an attempt to improve performance of the artificial neural network architectures being used today in various applications. Unfortunately there is no proper science to explain exactly what happens, or how, when we modify an artificial neural architecture, or combine it with other types of data processing architectures, for example.
The result of some of these modifications has been what are now being referred to loosely as “convolutional” neural networks, CNN (which include convolutional filters, for example); recurrent neural networks, RNN (which include feedback loops); and more recently, other types of variations including long short-term memory networks (LSTM); and perhaps most recently, a newer approach known as ODE networks. The list goes on, as new elements, which have been successfully used in other fields, find their way into the milieu of evolving neural networks. It is hoped that, with modifications, their performance will be improved enough that they might be credibly used in commercial or social applications. Let us remember that this is an empirical activity and very little is based on any kind of certainty or confidence that a result will ever be useful, or even the extent to which it might be useful.
At the present time, there can be so many possible variations on how an artificial neural network can be put together that finding or designing one that might be suitable for a particular application can be a major challenge. The cost to try to solve this problem may not be worth the human time and effort it would require. A truly suitable result might never be found, given the current state of the art.
Rube Goldberg Machines (RGM) and their Analogies with Current NN Research in AI – the RGM versus the Black Box
With the foregoing comments as introduction, and to perhaps clarify some salient features of AI today, we’d like to briefly suggest some analogies between these current efforts in artificial intelligence and the machinations of the mind of Rube Goldberg who was an engineer and popular cartoonist in his day (1883 – 1970).
In case you don’t happen to be familiar with the cartoons of Rube Goldberg, here are a few words of introduction. Perhaps it suffices to say that he devised and illustrated, in his comical drawings, often complex and humorous automated approaches (RGM) to performing simple or routine tasks which humans often perform ( but usually with significantly less effort}. To accomplish similar goals with computers is a primary goal of AI. To illustrate what we are talking about, here is an example of an RGM, showing how to automate a napkin for wiping one’s face while eating a meal (source – Wikipedia).
Professor Butts and the Self-Operating Napkin (1931). Soup spoon (A) is raised to mouth pulling string (B) and thereby jerking ladle (C), which throws cracker (D) past parrot (E). Parrot jumps after cracker and perch (F) tilts, upsetting seeds (G) into pail (H). Extra weight in pail pulls cord (I), which opens and ignites lighter (J), setting off skyrocket (K), which causes sickle (L) to cut string (M), allowing pendulum with attached napkin to swing back and forth, thereby wiping chin.
This cartoon presents in one simple image a number of analogies with current efforts in artificial intelligence which deserve to be commented on for the benefit of readers hoping to learn more about what artificial intelligence is all about.
To begin, we might note that a typical Rube Goldberg machine (RGM) only does one particular thing very well, similar to the situation in AI today. In the example above, it operates a napkin for one’s face while eating. Similarly, as in our present discussion, we see that the machine utilizes a number of components of various kinds which have been cobbled together in such a way that they are supposed to collectively accomplish the desired goal. To the extent that they did not do this, we may be sure the Rube would have found more components to throw into the mix until it worked as intended.
We may also note that this is an example of a feed-forward network, as is the basic NN described in our first Dialogue. Each successive component in the RGM acts to correctly activate the next component in the sequence from beginning to end (input to final output). The successive components in the RGM might be thought of as analogous to the successive layers in a basic artificial neural network, for example. In that case, each layer is just a single-layer NN, receiving an input and producing an output analogous to the components in the RGM.
In addition, if we are contemplating using this RGM “algorithm:, we need to note that it contains a particular component over which we may have very little control. That part is the parrot (E). It is not uncommon for Rube Goldberg to have included in his “algorithms” domestic (or other) animals who happen to work in ways that are not always predictable. For example, the parrot is expected to always catch a cracker (D) whenever it is thrown to him. There can be various reasons why he might fail to respond correctly to a cracker (D), including the possibility of not recognizing it as food. In any event, he represents an element of the algorithm which may behave in unpredictable ways that might just cause the desired result of the operation to fail.
We might then be justified in comparing the actions of the parrot to the action of a neural network which can also be unpredictable, and is poorly understood. In addition, in order to keep the algorithm functioning, a ready supply of crackers (D) must always be available in the ladle (C), for the algorithm to function. If we put any input other than a cracker into the ladle, the different input may just fail to evoke a correct response from the parrot, even if it might resemble a cracker in some ways. It might even be a cracker which is not exactly the same in appearance as the crackers which were used for training the parrot. The result again is that the algorithm can fail to produce the desired result, possibly due to just a minor variation in an input to the RGM.
There are, of course a number of other possible points of failure in the algorithm, including a lighter that might not light, and a rocket that might just have a faulty fuse or otherwise fail to ignite. Even if the fuse is good, if it happens to have been placed incorrectly into the machine so that the lighter cannot properly act on it, the machine will fail to accomplish its purpose. These represent components in the system that might not do what we expect them to do at the time or manner that we expect them to do it. These might correspond to components such as convolutional filters, added feedback loops, etc, whose actual operation when used within a neural network architecture might not be properly understood by us. Even the artificial NN architectures themselves are not well understood. In fact, one great advantage of the RGM is that we can understand how everything works and we can see where and why it fails, if that should occur (it is not a black box). We do not have this advantage when faced with the largely not-understood machinations of an artificial neural network (black box) of almost any kind, as is the present state of the art.
Looking Ahead – What’s Next?
Hopefully this brief discussion, with Rube Goldberg’s help and his comic cartoons, has shed some easily understood light on characteristic features of neural network-based approaches to AI today. The goal was to be able to examine some salient characteristics of efforts in AI today by way of a simple and easily understood analogy. If AI efforts today could be as easily understood and explained, there would be significantly less confusion about what we are dealing with and its implications.
This opens the door to the next post in which we’ll examine these issues further, and move ever closer to an understanding and conclusion to our original question (is it deja vu all over again?), and perhaps even a few more, as well.
Until then, thanks for dropping by and we’ll hope to see you again soon!