Lost in Translation – statistical learning isn’t panacea

Lost in Translation – statistical learning isn’t panacea

BBC Future published an excellent feature article on the status quo of machine translation.

Another good example of our naiveness, we think we can eventually achieve our goal – bringing peoples closer – through technology. For centuries we’ve been working on that, telephones, Skype, FB, and now Google translation. History has proved the futile nature of these kind attempts. The sarcastic tone at the end, originated from the Hitchhiker’s Guide to the Galaxy, sounds well grounded.

What the BBC article focuses on is a criticism of the current approach of statistical learning and its challenges it raises, including the vicious loop Google may end up falling into when using massive web text to train its translator: that it might be learning on its own translation (inaccurate and unnatural, shown by our experience, as well as the now infamous Malaysian website).

Natural language – modeling and manipulation (including translation and understanding) was a hot topic in the early days of AI, and has regained its status since the advances on statistical learning – and I argue it will probably remain a challenging problem deep into an unforeseeable future. Perhaps a real breakthrough can only occur when data such as our speech and pictures are truly conceptually grasped by a computer, when it not only does numbers crunching, logical reasoning, but also understands emotions – they will need to have emotions so as to be empathetic?

Do we really need that colossal amount of information – as much as Google now has – to master a language, or two? Is it possible to build a life-long learning machine, much like an infant (who grows into maturity)? Our speech is shaped by memory for sure, but has its own dynamics and uncertainties. Will a largely chaotic system that combines memory, pattern recognition, reasoning, and more “mental” processes together, be able to make a better translator?

Machine learning has come a long way, but the dawn is still not there yet.

So many faces…

Tags

face, saliency

Spotting a face out from a cluttered background may be an effortless thing, but have you ever tried searching out a friend’s face in a group photo, and you wondered why it took you so long?

In this Face Facts: People Don’t Stand Out In A Crowd article, this crowding effect was first studied for high-level stimuli such as faces. More details are available from this JoV paper by Bressler and Whitney.

What does this relate to our research in modeling high-level saliency in perception? With many saliency points contending our focus, surely the saliency level of these points will cancel each other, perhaps using some inhibition mechanism.

On the other hand, can we find a signature space for faces, so that a face hunt among a crowd can speed up, perhaps using a multi-resolution approach, so that identification (and localization) of a certain face can be more efficient, without going through scanning through all the faces one after another on the finest resolution?

Contrary to the notion that a face is processed as an image in our vision system, this JoV paper treats faces mainly as horizontal lines or bar-codes!

Faces are interesting, aren’t they?

The role of visual attention – it needs a balanced view?

The Science Daily reports two research findings slightly in ‘contradiction’ with each other.

On one hand, visual attention is always important in perception. Our visual system has limited capacity and can’t afford processing everything fallen into our sight. A need for focus suggests that parallel processing or pre-processing in retina or visual cortex do not point to final perception directly, which has to be processed using attention-based mechanism, with eye movement and tracing involved. A recent article discusses the role of centre-surround receptive fields in an attentional system.

On the other, viewers can learn a lot about environment even without paying attention – see this article. Staring at a very familiar scene may not produce any mental attention – I suspect visual focus is still on target, but the mental process can simply block the stimulus. Subconsciously, it is other information that gets processed. Although this makes perception a more complicated issue, arguably it’s a good thing for us. Many times it occurred to me, that I was watching things going on the street, but all of a sudden, I stop just inches away from something protruding to my face – a twig, or a lamp post. Without attention, my side view is working silently, and that is crucially important! In a sense we see things that catch our intention – even though the focusing might be carried out subconsciously (see this example which is also quite relevant, and fun).

So, how much attention should we pay to the research of visual attention )

Associative memory study – time for a change?

Tags

memory, associative, pattern, neuroscience

How do we remember faces and other patterns?
I’ve always suspected that what we remember is not the pattern per se, but rather a kind of feeling uniquely connected to that pattern. So, seeing a somewhat familiar face, we don’t really try to match it with a model template, pixel-by-pixel, stored in our brain somewhere, but rather, it’s the same perception (sometimes can be emotional) response that repeats itself, something similar to a deja vu (however with stronger intensity), which gives us the certainty about the identity of the face.
That, may be extended to the memory of landscape, drawings, and music.
For decades, computer scientists have been working on building associative memory models using artificial neural network models. No matter it is auto-associative, hetero-associative, bidirectional memory, or adding a bit dynamics – chaotic associative memory, the computational model is almost always built on the Hebb rule, using correlation matrix to store the spatial patterns per se.
Neuroscientists have long found the relevance of eye movement and saliency detection in object recognition. In light of the recent neuroscience finding, The Brain ‘Joins The Dots’ When Drawing A Cartoon Face From Memory, it becomes clearer that we should build our memory models otherwise, incorporating key points and their tracking configured in an ‘action plan’. In other words, not only is spatial correlation important, but also the timing of when attention shifts. Or it could be the sequence of firing neurons along the timeline, and its pattern, that matters? Therefore we have to build large-scale parallel systems so as to cope with the memory formation and retrieval… but how?
Well, I don’t really see an answer to these questions. Do they make sense to you?