I for one really liked the demo and the blog - specifically, (a) I have great exemplars for what you mean by "theme", and (b) this post[1] shows great insights into your thinking about the problem faced by your customers
> Developed on canonical text like news article or Wikipedia, they either failed to understand the variety of expressions, or were too hard to explain.
It appears to me that the current methods and resulting tools are heavily dependent on the problem formulation (or domain in general). Moreover, no matter how fancy your technique is (or "how deep is your net"), the resulting model won't work unless you take specific steps to train it on data from the domain.
Yes, what I just said sounds borderline truism. However, I am more interested in discussing why it is so? Here's my initial thinking:
Let us look at (one of) the definition of Machine Learning, from Prof Tom Mitchell's textbook,
"A computer program is said to learn from experience E with respect to some class of tasks T and performance measure P, if its performance at tasks in T, as measured by P, improves with experience E."
Here, experience E can be loosely considered as the amount of data you have for training - obviously, more data (i.e. training) should improve learning. However, the abstraction of T and P hides an important underlying problem of specification - or in other words, formulation of T (and E).
Thoughts?
> I wrote a new approach [capitalizing] on my PhD and new Deep Learning approaches.
I hope we get to see some of your insights in a paper or article (or blog post :)
Thanks for your thoughts! Definitely sounds like an interesting thought to explore.
I like to think about ML in terms of how children learn language: through observation in their environment. The training data is a simulation of that environment.
I for one really liked the demo and the blog - specifically, (a) I have great exemplars for what you mean by "theme", and (b) this post[1] shows great insights into your thinking about the problem faced by your customers
> Developed on canonical text like news article or Wikipedia, they either failed to understand the variety of expressions, or were too hard to explain.
It appears to me that the current methods and resulting tools are heavily dependent on the problem formulation (or domain in general). Moreover, no matter how fancy your technique is (or "how deep is your net"), the resulting model won't work unless you take specific steps to train it on data from the domain.
Yes, what I just said sounds borderline truism. However, I am more interested in discussing why it is so? Here's my initial thinking:
Let us look at (one of) the definition of Machine Learning, from Prof Tom Mitchell's textbook, "A computer program is said to learn from experience E with respect to some class of tasks T and performance measure P, if its performance at tasks in T, as measured by P, improves with experience E."
Here, experience E can be loosely considered as the amount of data you have for training - obviously, more data (i.e. training) should improve learning. However, the abstraction of T and P hides an important underlying problem of specification - or in other words, formulation of T (and E).
Thoughts?
> I wrote a new approach [capitalizing] on my PhD and new Deep Learning approaches.
I hope we get to see some of your insights in a paper or article (or blog post :)
[1] https://www.getthematic.com/post/visualizing-customer-feedba...