It’s all in your Mind
By Tetyana Loskutova
Today Big Data equals big hype. It might be a strange idea at first, since the data was there from the beginning of computing; storage and processing of data was, essentially, the main operation of computers since Turing Machine.
However, the problem is in the name. Essential characteristics (or essential problems) of Big Data is not just it’s size, it is its velocity and variety.
Big Data is essentially real world. Computer challenged with processing of big data faces the same problem as we, humans, face at each moment of our consciousness.
The moment we wake up, our brain is bombarded with ever-changing, large variety sensory inputs: sound of the alarm clock, temperature of the air, level of light, smell of coffee, hunger, memories of the things to do etc.
This highly variable input keeps changing every moment throughout the day and our brain has to quickly classify and timeously process this data and initiate output reaction.
How similar it is to a data centre that processes LinkedIn feed? The data comes as text, images, CVs, videos, i.e. similar to the different senses in the body, and it changes all the time in a non-linear fashion where knowledge of the past does not guarantee prediction of the future. Just like human sensory data, LinkedIn feed data may be linked (like comments), related (different posts from the same location), relevant to the observer or not, charged with different emotions, metaphors and symbolic meanings. These kind of meanings require understanding. Can machines understand?
According to Jeff Hawkins, neuroscientist and creator of Palm Pilot, this can be achieved by modelling the work of human brain – neocortex. The history of computing was shaped by the tasks of calculation based on certain logical rules, which were pre-programmed. Though the idea of self-learning machines and artificial intelligence was around for a good 50 years, the results up to this day are still lacking the notion of machine intellect, cognition and understanding. Computers can do complex calculations in nano-seconds, but have difficulty distinguishing images of different animals – a task that 6-year-old child can do in seconds.
Long quantitative repetitive predictive tasks are easy for machines, that’s the territory where they outperform humans with ease. In the same time processing of large qualitative data is endlessly challenging. Currently used techniques of word search, word count, pattern recognition help to structure and reduce the data, however Big Data is always changing – adding new topics, new words, new expressions and new symbolic meanings. In order to effectively deal with this type of data, machines need to learn and evolve with the data – using intellect similar to the human one.
Artificial Intelligence supporters are optimistic that addition of computing power will make this task possible some day. However, it is possible to prove that human brain can already fulfil the tasks of understanding using much smaller computing power. Artificial Intelligence may be intelligence (depends on how one defines it), but it is indeed very different from ours.
Brain is not a computing device, it is an ultimate pattern recognition device. Differently to computers, human brain uses patterns for learning and prediction. When learning foreign language students can often catch the meaning of the conversation without understanding all the words. When these words are repeated in different settings, students will eventually learn the meaning of the unknown words without translation. This ability can be attributed to several important structural features of neocortex: hierarchical network organisation of neurons, hierarchical memory and storage of invariant patterns.
Hierarchical network of neurons has inspired creation of neural networks – artificial network structures imitating the organisation of neocortex. Even the simplest 3 level artificial neural network can learn the pattern and then return the full pattern when given only apart of it. The same process happens in the brain when we recognise the song from the first few notes. However, brain’s structure is much more complex. Patterns are stored sequentially and retrieved auto-associatively. We never remember the whole route in the city, but starting to drive, we can always retrieve the memories of the next turn once we approach it. The same happens to older memories – we may never recall that we’ve been to a certain place, but once we see a picture of it, we can describe in detail what we did there.
This auto-associative information retrieval is much faster that the full memory scan or full re-calculation routinely used in computer industry. Another unique – and possibly most difficult – ability of human brain is the storage of memories in their invariant form. This is the ability that helps us to recognise dogs from other animals independently on which breed of a dog it is.
Our brain stores some pattern that is associated with the definition of a dog, and variance from this pattern, occurring in reality, is acceptable, predictable and learn-able. Once you saw a dog that looks very similar to a cat, your brain will store additional pattern that have helped you to make sure that the dog is not a cat ;and this pattern will get associated with the definition of a dog. Next time you see that type of a dog – you will know immediately.
This seemingly simple algorithm has proven to be endlessly difficult to program, because current computers are not very suitable to reduce objects to ever varying invariant patterns – even the definition seems illogical.
Brain is uniquely suitable for analysing unstructured ever varying data. Effectively, the brain is the best tool to analyse Big Data when scaled to the size suitable for the data stream. Brain-like artificial networks have the potential to process ever changing Big Data without losing effectiveness over time. However, with the brain-like intelligence – will we have brain-like errors?
1. “Machine Learning Attacks Against the Asirra CAPTCHA”
2. Jeff Hawkins “On Intelligence”
3. Open source Numenta Platform for intelligent computing contains a set of learning algorythms and hierarchical temporal memory.