Combine multiple capabilities of present day computers, and you very well may be able to piece history back together.
How is this possible? It'll take several tools, but the most amazing thing is that we have these tools. The most recent one (as of 10 September, 2016) is the Scene Dynamics algorithm.
Conditional Video Generations
We can also use the model to add animations to static images by training it to generate the future. Of course, the future is uncertain, so the model rarely generates the "correct" future, but we think the prediction has some plausibility.
Sentence and paragraph vectors are also advancing at an astounding rate, and we've just seen how DeepMind can synthesize realistic audio.
This gave me an idea.
Right now, the algorithm is very crude and can't animate things very well— hence why it's back to 'dogslugs' ala DeepDreams. But that's because it's putting imagines into motion based on what it 'thinks' will happen next.
But what if it knew what happened next? At least, what happened a bit further into the future? What if you gave it a picture, and then gave it another picture that was taken about 5 minutes later, and told it to 'fill in the blanks'?
Just by using these two pictures, you might get a somewhat accurate reconstruction of history. However, there's more to history than just the direct events that happen. The accuracy of the image can be increased exponentially if you factor in external events. For example— in Picture A, you might have the subject matter be 2 men with a whole bunch of people in the background. In Picture B, you only have 1 man, and while you still have a lot of background people, there's clearly a new set of them. The algorithm can make a guess of what happened, and if you give it a narrative, it should be able to create a general video of events.
Now take in a whole bunch of seemingly irrelevant information, such as Facebook updates, news stories, etc. The resulting video might not seem much different at first, until you compare what goes on in it to what actually happened that day. All of a sudden, what was maybe a 50% accuracy rate became an 80% accuracy, because now you know the reasoning behind some of these people's movements, why they were in this or that particular picture, etc.
Now take all this and add one final point— the present. Feed the algorithm information about the present, and see if it can build a timeline of events on how it got there from the two photographs. If something doesn't add up— whether it's something actively happening in the picture or something happening in the background, well out of the mind of any of the subjects— the algorithm will be able to tell.
It still won't be perfect since there are simply so many variables, but it could go a long way towards helping us figure out how history unfolded.