Jump to content

Artificial Intelligence


Parsad

Recommended Posts

I thought I would start this thread on the field of AI which is garnering more and more investment, and is starting to affect our day to day lives.  The area will be greatly beneficial to humanity, but carries certain risks, and could be extremely disruptive to many businesses.  Feel free to post articles on here and carry on discussion regarding AI.  Cheers!

 

Link to comment
Share on other sites

Automatic decision making is the future (UPST, FICO, etc.). Decisions and choices are influenced by the media, friends, etc., usually without people being aware of that the decision or choice they thought they made was not their own (😒 $DJCO 13F-HR). In the future everything will be automated even more, Zero-Click Ordering ($DPZ), etc.:

 

 

 

Edited by formthirteen
Link to comment
Share on other sites

7 hours ago, formthirteen said:

Automatic decision making is the future (UPST, FICO, etc.).

 

This is why Weapons of Math Destructions by Cathy O'Neil has been a mandatory read for students in my class since the day the book was published. It's an easy read that gives a glimpse what bad automation can look like.

 

I do agree that AI will permeate every aspect of our lives. What I find interesting is AI is entrenching in low and high value tasks almost simultaneously. Low end of the spectrum makes money on volume (think forms processing). High end makes $ on complexity (think Tesla CV system). Ends users are generally oblivious to what is happening and user adoption is consistently high. Implementation costs is what varies.

 

On a personal note, I'm very well versed in graph and NLP spaces. I can't wait for Neo4j to IPO. Further maturation of NLP will be big too (moving beyond BERT and T5).

Link to comment
Share on other sites

I have been self-learning ML and deep learning for many months, Im at the beginning of the journey but it's a really fascinating field. My futuristic stand point is that we are going to automatize every boring process of our life freeing space for our creativity, the real superpower of human beings. Decision making, as suggested by Kahneman, will be left to the machine and in my opinion there will be some sort of connection with a global and immutable blockchain turing complete.

I wonder how AI will change the investing environment.

Edited by Dave86ch
Link to comment
Share on other sites

There are thousands of machine learning algorithms out there, but you'll rarely need more than a handful.

A good start:

• Linear/Logistic Regression

• Decision Trees

• Neural Networks

• XGBoost

• Naive Bayes

• PCA

• KNN

• Random Forests

• K-Means

Link to comment
Share on other sites

52 minutes ago, drzola said:

There are thousands of machine learning algorithms out there, but you'll rarely need more than a handful.

A good start:

• Linear/Logistic Regression

• Decision Trees

• Neural Networks

• XGBoost

• Naive Bayes

• PCA

• KNN

• Random Forests

• K-Means

Some of these are algorithms but some of these are whole domains, e.g., Neural Networks, and on their own won't get you very far. Most of the algos listed are a good theoretical/academic start but have been proven to work well in a limited set of domains. To further complicate things, any respectable shop will have pipelines with 10,000+ models at any given time.

 

The recent R&D from NVIDIA, MSFT, Google take neural networks (GANs) to  entirely new level. Just take a look at NVIDIA Riva SDK (anyone remembers Riva GPU from the 90s!; the two are not related other than both being from NVIDIA). Google's BERT and T5 fall into the same category. There is a massive push to combine AI with more domain specific knowledge (think spectral analysis for voice, etc.) to get good results. 

Link to comment
Share on other sites

A Paperclip Maximizer is a hypothetical artificial intelligence whose utility function values something that humans would consider almost worthless, like maximizing the number of paperclips in the universe. The paperclip maximizer is the canonical thought experiment showing how an artificial general intelligence, even one designed competently and without malice, could ultimately destroy humanity. The thought experiment shows that AIs with apparently innocuous values could pose an existential threat.

 

The goal of maximizing paperclips is chosen for illustrative purposes because it is very unlikely to be implemented, and has little apparent danger or emotional load (in contrast to, for example, curing cancer or winning wars). This produces a thought experiment which shows the contingency of human values: An extremely powerful optimizer (a highly intelligent agent) could seek goals that are completely alien to ours (orthogonality thesis), and as a side-effect destroy us by consuming resources essential to our survival.

 

www.lesswrong.com/tag/paperclip-maximizer

Link to comment
Share on other sites

Cant run AI without hardware.


NVDA has the lead on ANN training market. Recent ARM purchase attempt was to capture ANN inferencing market share. If this fails, Jensen will have his way at all cost, probably an  in house CPU from NVDA will be in order.

 

Intel is coming at it from both directions. Pote Vecchio for high throughput workloads for ANN training, and AMX instruction set for matrix operations on their Sapphire Rapids server CPU to go after inferencing market share.

 

The up and coming inferencing contender is probably Cerebras claiming low latency( inferencing) on an accelerator ( training) platform.

 

But according to Naveen Rao (podcast) founder of nervana ( acquired by Intel), the blurring of training and inference, although is the ultimate future, is yet a few decades away, and credits NVDA as defacto winner of AI hardware of today (who knows better than outsiders how easily the lead could be toppled, and is not standing still).

 

I am really interested to see how the Neuromorphic computing platform by Intel plays out. ( currently available for tinkering via intel cloud). Today’s software paradigms are unfit for programming anything on here. Ultra low frequency, ultra low power, asynchronous computing; hardware that mirrors its algorithms.

Edited by Minseok
Link to comment
Share on other sites

On 12/1/2021 at 2:49 PM, Dave86ch said:

I have been self-learning ML and deep learning for many months, Im at the beginning of the journey but it's a really fascinating field. My futuristic stand point is that we are going to automatize every boring process of our life freeing space for our creativity, the real superpower of human beings. Decision making, as suggested by Kahneman, will be left to the machine and in my opinion there will be some sort of connection with a global and immutable blockchain turing complete.

I wonder how AI will change the investing environment.

 

As someone who is fairly well versed in both investing and the AI field, I am fairly skeptical it will gain much of a toehold in investing. There are two main problems in the application of AI to investing:

 

i) Data Drift - The investing world changes over time, and the drivers of the market today are likely much different than they were in historical data. Most of the AI technology you see hyped in the media works because it has access to huge datasets which are relatively static (i.e. pictures of cats and dogs). Many of these algorithms do not translate into something like investing which have relatively small and changing datasets. 

 

ii) Cost of Acquiring Data - In many AI algorithms the variables used to predict an outcome are easily available. For example, in determining whether a picture is a cat or a dog from an image, you have the pixels of an image, and that should be all you need to make a prediction. In investing it is different. For example, maybe you think that a good predictor of returns going forward is normalized operating cash flow per share divided by the price of the stock. So if you want to train a model, with lets say 100 stocks, you need to go an calculate the normalized operating cash flow for each stock for each period going back many years. But it might be nontrivial to find out what these normalized values are (maybe there were idiosyncratic factors, or maybe firms made acquisitions, or maybe a specific sector was in a slump but not anymore, maybe there were regulatory changes). There is a saying in ML 'garbage in garbage out' which basically means if you fail to properly prepare your dataset your predictions will be worthless. Because it is so time consuming to get 'clean' data for a long period of time, and all of the difficulties associated with it, it is unclear how much one's model would outperform a more 'traditional finance approach' to investing.

 

I think if you are interested in learning more about quantitative methods that could be applied to investing you stick with statistics and more basic machine learning methods (i.e. not deep learning).

Link to comment
Share on other sites

2 hours ago, maplevalue said:

 

As someone who is fairly well versed in both investing and the AI field, I am fairly skeptical it will gain much of a toehold in investing. There are two main problems in the application of AI to investing:

 

i) Data Drift - The investing world changes over time, and the drivers of the market today are likely much different than they were in historical data. Most of the AI technology you see hyped in the media works because it has access to huge datasets which are relatively static (i.e. pictures of cats and dogs). Many of these algorithms do not translate into something like investing which have relatively small and changing datasets. 

 

ii) Cost of Acquiring Data - In many AI algorithms the variables used to predict an outcome are easily available. For example, in determining whether a picture is a cat or a dog from an image, you have the pixels of an image, and that should be all you need to make a prediction. In investing it is different. For example, maybe you think that a good predictor of returns going forward is normalized operating cash flow per share divided by the price of the stock. So if you want to train a model, with lets say 100 stocks, you need to go an calculate the normalized operating cash flow for each stock for each period going back many years. But it might be nontrivial to find out what these normalized values are (maybe there were idiosyncratic factors, or maybe firms made acquisitions, or maybe a specific sector was in a slump but not anymore, maybe there were regulatory changes). There is a saying in ML 'garbage in garbage out' which basically means if you fail to properly prepare your dataset your predictions will be worthless. Because it is so time consuming to get 'clean' data for a long period of time, and all of the difficulties associated with it, it is unclear how much one's model would outperform a more 'traditional finance approach' to investing.

 

I think if you are interested in learning more about quantitative methods that could be applied to investing you stick with statistics and more basic machine learning methods (i.e. not deep learning).

I can see where you coming from, I noticed this current "rigidity" of the models that doesnt fit well with a complex adaptive system. Anyway this field is at the beginning of its journey, surely there will be a lot of surprises.

Link to comment
Share on other sites

8 hours ago, drzola said:

interesting here inquiry of;  The conflict of Laws of AI usage and TORT

 

https://repository.law.miami.edu/cgi/viewcontent.cgi?article=1678&context=fac_articles

This thread and specific post are interesting.

In a relevant way for 'diagnostic analytics', there was a publication showing superiority of the model output over human output for a specific task: from a picture, decide if the skin lesion was benign or malignant. After peer-review (humans still in the loop), it was discovered that the computer 'cheated' as it associated the presence of a ruler in the picture with a higher risk for malignancy and this was simply indicating a certain level of the concern by the human who tended to measure (with a ruler) the lesion when the human somehow 'felt' that the lesion had malignant characteristics. So, once this 'bias' was removed, the superiority of the model decreased significantly. A fascinating aspect is that the multi-level algorithms will tend to integrate implicit human biases when mixed within the supposedly 'raw' data.

This area is still in infancy but the potential is huge.

It's likely that the best outcome will be a variable hybrid 'system'. The origin of AI's output can be well traced but poorly explained (black-box 'reasoning') whereas the origin of human brain's output can be well explained (not always..) but poorly traced. This aspect is also relevant for investing i guess.

Personal and anecdotal comment: since graduation, there's been this huge (and welcome) movement to improve the human side of the (diagnostic and treatment) conversation and now 'machines' are  about to take over the world..

 

Link to comment
Share on other sites

Create an account or sign in to comment

You need to be a member in order to leave a comment

Create an account

Sign up for a new account in our community. It's easy!

Register a new account

Sign in

Already have an account? Sign in here.

Sign In Now
×
×
  • Create New...