INSPIRED BY BRAIN!
Chandrakant Pattekar
Cloud & Core - Large Deliveries, Portfolio management, Strategic Initiatives
Much of the research on Artificial Intelligence has been inspired by the brain and its functioning. Perceptrons or some prefer to call it the sigmoid neurons that are the granular software entities of neural networks emulate the neurons inside the brain. Akin to the dendritic tree of the neurons, the perceptrons receive inputs from many other perceptrons. The receiving perceptron processes the weighted sum of these inputs to produce an output which is then relayed in a similar way to another bunch of perceptrons. This web of preceptrons forms the basis for advanced neural networks. I guess this only a small part of the overall inspiration that AI stands to draw from brain in a longer run.
Some time back, I was at the brain museum in Bangalore with the curiosity of a 10 year old. The tour included an overview session on brain anatomy that lasted about an hour. There, I experienced one of the most amazing moments when I held the 1.2 kg organ in my bare hands. Many mysteries lay embedded within it. Every living organism is endowed with a brain however tiny or large it may be. It compels oneself to wonder about how our modern age electronic computing underpinning the AI compares with this million years old computing machine created by the Great Master.
There is one fundamental difference between artificial neural networks and the neuronal web inside the brain. Each neuron that’s part of the 85 billion neurons inside the brain is a living entity. Each neuron consumes oxygen and energy for its functioning. This gives an independent ability to every neuron to process the incoming signals. As opposed to this the perceptrons are essentially software entities and the real processing happens in the CPU or GPU cores. The GPUs have an edge over the CPUs as it has many hundreds of cores that can make ANNs work many times faster. But this hardly compares with the 85 billion neurons in brain. The Tensor Processing Units being developed by Google may bring further improvements in this regard. Intel is also working on faster processing units to support AI. May be we shall soon see advent of new generation microscopic processing units that will bridge this gap that exists today.
Till very recently, our machines were blind, deaf and dumb. This has changed with AI. Let us take a closer look at the visual processing capabilities in AI. When trained, CNNs can classify images and can also label the actions seen in videos. But we still have a lot of ground to cover. ‘Image classification’ needs to evolve and lead to ‘object recognition’ as AI matures further. As an example, let us say that a CNN classifies a particular image captured by the camera of a mobile phone as a ‘ball’ correctly. To make this more meaningful CNN needs to see the ball as an object rather than just a label. In other words, CNN should be able to recognize various attributes of the object along with the actions that can be performed with the object or by the object. E.g. in this case one can throw, catch, bounce or kick a ball. Also that the given ball is soft, round, small, smooth etc. This will take AI to the next level where machines can learn about the possible ways to interact with a given object in a particular situation. Having said this, let us not forget that artificial intelligence is not about hand programming but is about developing a learning ability. It may be easy to hand program the attributes and capabilities of a large number of objects as in OOP but this not the intent as it will never produce a scalable solution that AI possibly can.
Similarly, 3D CNN can identify the actions seen in a video clip and can classify the actions by labeling them. For example, a CNN network can be trained to recognize actions such as walking, talking, sleeping etc. as seen in a given video clip. This needs to evolve further such that AI can predict outcomes resulting from the action in progress. This will make our machines smarter to suggest that one needs to ‘walk’ up to the ‘ball’ to ‘kick’ it.
In summary, AI needs to find a way for the machines to learn the following:
· Object recognition inclusive of object attributes and object capabilities
· Action recognition with ability to predict potential outcomes from a given action
With the above, as if gamifying, the machines would engage in a constant interplay of objects and actions to create hypothesized outcomes. These hypothesized outcomes may be right or wrong, likely or unlikely – it does not matter. The machine learning shall entail comparing these hypothesized outcomes with the real life situations. The learning gives higher weightage to the right or likely outcomes than the ones that are wrong or unlikely thus enabling the machine to predict an action plan that is based on learning from real experience. Surely, this is just a high level discussion and the devil lies in the details.
This will take us a little bit closer to one of the end goals for AI – to develop the next generation machines that can be trained to predict an action plan and execute it in a given situation. Isn’t that something that we do in every living moment of our life?
Head of Delivery Strategy at Ericsson India Global Services Pvt Ltd
6 年Brilliant article...I have an article to share on similar topic? https://www.dhirubhai.net/pulse/humans-more-intelligent-than-computers-sandeep/
Head of Delivery Strategy at Ericsson India Global Services Pvt Ltd
6 年Humans are more intelligent than computers The rising power of computers and advances in Artificial Intelligence has reenergized the debate of intelligence of computers relative to humans. Broadly, there are two approaches to it. First approach focusses on study of human brains and compares that with computers. Human brain has about 100 billion neurons which are often compared to gates in computers. Neurons interconnect through synapses and may have upto 10000 connections each giving a total of 100 to 1000 trillion connections. Neurons communicate through electrochemical signals which are one million times slower than speed of signals on fibre optic cables and neurons fire at about only 200 times a second. They collaborate to provide about 2.5 petabytes of memory. Brain consumes about 20 W of power. Supercomputers have similar or better numbers except that they need power in MW and most importantly the transistors connect with very small number of neighbors and have 2D geometry. But computers cannot be compared with human brain on these numbers alone due as human brain is not a digital machine driven by a clock and it does not differentiate memory and processing areas. Besides, the functioning of brain is itself not properly understood. A simulation of human brain by Fujitsu-built K took about 40 minutes to complete simulation of one second of neuronal network activity in real time in 2013. Second and the more popular approach focusses on imitating the functionality of the brain. This approach also does away with the difficult question of defining intelligence. Computers are definitely ahead of humans in calculations or when it comes to executing simple step-by-step instructions. This power was at display in 1997 when IBM’s Deep Blue, a computer defeated the then world chess champion, Garry Kasparov. The computer could evaluate millions of possible positions per second and think of the next 20 moves. And today super computers with speed in peta flops (10^15) can outwit any human in any calculative task. One important quality of humans is that they can learn. Machine learning is the branch of Artificial Intelligence which focuses on creating machines or computers that can “learn”. So, driverless vehicles, email filtering, spam detection and most importantly robotics is based on this idea. And success of Google’s AlphaGo in defeating champion, Lee Sedol in game, Go is based on Machine Learning as number of possible positions in Go is so large that even fastest computers would be swamped. But computers need to be programmed or told on how to “learn” and computers thus programmed will work only for those situations. Besides, success in these areas does not translate into superiority of computers e.g. while driverless vehicles will “learn”, they still need detailed 3D maps to work efficiently unlike humans. One important attribute of human brains is the ability to recognize patterns e.g. characters, faces, voice in noise. This is an area where computers do not match human abilities. Computers can recognize printed letters and numbers, and can recognize specific faces and automatically tag photos of those people as you take pictures. But humans can recognize complex patterns and adapt to them. Humans can also recognize faces which are covered with facial hair, have done make ups etc. A technique of Machine Learning called Deep Learning is used to train computers on pattern recognition. Here it creates layers of nodes with interconnection between layers similar to that of neurons in the brain. Same applies for language abilities of human. As of now, computers can do simple translations between 2 languages, speech to text translation and vice versa. Again, deep learning is being used to improve capabilities of computers. IBM’s Watson which won in Jeopardy! In 2011 had to use Natural language processing as the game needs questions to be created against answers. But computers are still behind as human languages are ambiguous and the linguistic structure can depend on many complex variables like slang, regional dialects etc Computers need to bridge the gap on creativity front. Though there has been some progress here too. Computers are being used for writing news in Washington Posts, USA Today, Wired etc. Shimon, a robot from Georgia Institute of Technology can compose music and there are competitions to display paintings created by robots. However, human’s abilities in arts e.g. writing stories and poems, making paintings, composing music etc are beyond the reach of current set of computers. Same applies to research. Adam, a robot designed by British scientists is capable formulating hypotheses, designing and running experiments, analyzing data, and deciding which experiments to run next. But current set of computers cannot formulate new scientific theories. Computer engineers at Cornell University designed a program that could give a computer basic set of tools it could use to observe and analyze the movements of a pendulum. Using this foundation, the software was able to extrapolate basic laws of physics from the pendulum's motions. But the computer could not create the tools of its own. Additionally, the computers do not have feelings e.g. love, fear, anger etc. This fuels ambition and creativity and advances civilization. Superior feelings of humans allows them to make much stronger bonds with a much larger geographical spread and has contributed to their dominance over other species. This bonding allows humans to benefit from collective intelligence of mankind rather than an individual. Computers cannot reason or understand impact of a decision. Computers are yet to beat humans in a game called Startrek which requires lots of decision making. Turing test devised in 1950 checks if a computer can mislead a human to think that it is human based on a conversation, a sort of imitation test. In 2014 a Russian-designed programme called Eugene Goostman could mislead 33% of the judges in one Turing test. Since then other tests e.g. Loebner etc have been devised with more difficult criteria including higher misleading rate. Computers will be able to imitate more aspects of human brain in the next few decades but a computer capable to having all of abilities of human brain and equally efficient on resource utilization is still in realms of science fiction.
Certified Career Coach I Strategic Leadership | Operational Excellence | Merger Acquisition | D&I
8 年Very well written! Chandrakant Enjoyed reading it. ..
Product development leader
8 年Great article Chandrakant Pattekar. Would love to hear what challenges are there now in classifying the multiple attributes of an object - say texture, levels of granularity? Are those computational challenges or we do not have a model today for fine grained classification? Would love to hear your thoughts on what pushes different people to notice different things in different perspectives? What makes different people draw different conclusions / interpretations / inferences from same / similar set of data?
Head of Delivery - MEA Region
8 年Interesting article Chandrakant