1,510 karma · joined May 23, 2012
I really wish I knew how to resolve this problem, but it feels more like a physical problem, to do with the limits of human attention and learning, than a social one. I suspect there is no easy fix.
The idea has evolved over the last few years during the course of regular weekly discussions held by our group of founders.
Our group is planning on turning our technical developments into a startup, probably initially focusing on systems engineering tools.
We are currently implementing our prototype and looking at potential application areas.
It's horrible.
Our ML algorithms are really good at finding correlations -- but we don't necessarily know if the correlations in our data are actually the ones we want our system to learn. When we're using synthetic data, we have many more levers at our disposal to ensure that this is the case.
So, as a simple example, the illumination in a real data-set might be strongly bimodal, with comparatively few samples at dawn and dusk, but we might in a synthetic dataset want to sample light levels uniformly across a range that is specified in the requirements document.
Similarly, on the road, the majority of other vehicles are seen either head-on or tail-on, but we might want to sample uniformly over different target orientations to ensure that our performance is uniform, easily understood, and does not contain any gaps in coverage.
Similarly, operational experience might highlight certain scenarios as being a particularly high risk. We might want to over-sample in those areas as part of a safety strategy in which we use logging to identify near-miss or elevated-risk scenarios and then bolster our dataset in those areas.
In general, the synthetic dataset should cover the real distribution .. but you may want it to be larger than the real distribution and focus more on edge-cases which may not occur all that often but which either simplify things for your requirements specification, or provide extra safety assurance.
Also, given that it's impossible to make synthetic data that's exactly photo-realistic, you also want enough variation in enough different directions to ensure that you can generalize over the synthetic-to-real gap.
Also, I'm not sure how much sense the concepts of mean and variance make in these very very high dimensional spaces.
Because you have this piecewise-linear sort of warping of the feature space going on, the NN is basically a whole bunch of lever-arms. The broader the support that you can give those lever arms, the less they will be influenced by noise and randomness ... hence my obsession with putting enough variance into the dataset along relevant dimensions.
To put this another way, I think that the synthetic data manifold has to be 'fat' in all the right places.
To my knowledge, the best public attempt at this is represented by the OpenAI Rubik's Cube / Shadow Robotics dexterous hand demo. https://openai.com/blog/solving-rubiks-cube/ https://arxiv.org/abs/1910.07113
NVIDIA are also doing some interesting work in this area, but again, I'm not really sure they put enough different modes of variation into it. https://research.nvidia.com/publication/2018-04_Training-Dee....
CVEDIA also get really impressive results using similar techniques:- https://www.cvedia.com/
Of greater concern are quality measures that look across the entire dataset. Here are some hypothetical metrics which (although impossible to compute in practice) will help get you thinking in the right way.
- How does the synthetic image manifold compare to the natural image manifold?
- Are there any points on the synthetic image manifold where the local number of dimensions is significantly less than at the corresponding point on the natural image manifold? (Would indicate an inability to generalise across that particular mode of variation in that part of feature space).
- For each point on the synthetic image manifold, are there any points where the distance between the synthetic image manifold and the natural image manifold is large AND the variance of the synthetic image manifold in the direction of that difference is small. (Would indicate an inability to generalise across the synthetic-to-real gap at that point in the manifold).
- Does your synthetic data systematically capture the correlations that you wish your learning algorithm to learn?
- Does your synthetic data systematically eliminate the confounding correlations that may be present in nature but which do not necessarily indicate the presence of your target of interest.
Engineering with synthetic data is not data mining. It is much more akin to feature engineering.
In fact, it works so well it feels like cheating.