26-01-2022
Good luck from Zillow's success on his algorithms.Algorithms, worth $ 1 million. For the developer, the company benefited from the talent base with which it connects.
However, it was also an algorithm that caused the recent debacle of the company.Unpredictability of predicting house prices that are much larger than what we expected and that Zillow offers continue to scale, would lead to too much income and volatility for balance, "said Zillow CEO Rich Barton in a statement.$ 304 million for Q3 2021.
What went wrong?
The mistake of forgiving house prices can be attributed to a large number of possible causes, such as the response of the market to the macro -economic policy (ie unprecedented monetary exemption), unusual work conditions or the pandemic.Insecurity.Just like autumn of LTCM (long -term capital control) in the late 90s, the concern of people about AI algorithms are mainly due to (un) interpretability.In the LTCM case, people clearly knew (although they were a bit late) in 1998 he broke some mathematical assumptions about the quantitative strategy.Good.
Does interpretability really mean something?
We still cannot fully understand why Zillow's algorithm failed, just as we cannot fully understand why Alphago became a top player without knowing a pre-defined strategy, and many AI-Chatbots unexpectedly learn all prejudices in human society without design intervention.Only blame the nature of (in) interpretability and try to disrupt or even forbid the use of AI in many areas, such as healthcare, justice or social media.
However, this can be a pessimistic idea: compared to the human mind, inexplicable AI algorithms can be closer to how the world works.We can understand this problem with regard to the fundamental way in which deep neural networks work.and their variants).If we understand it in an abstract way, the activation function makes it easy to use basic operations to achieve something like a "conditional explanation": the neuron is "excited" or "advised" of the conditions that the input signal meets.In human neuronal cells, a similar process of the selective binding of receptor proteins on the post -synaptic membrane as neurotransmitters obtained.
Figure 1 Organic Neuron and his mathematical model
Why is this real estate with "conditional explanations" interesting?We can imagine a neural network as a baby who has just started exploring situations (movement, communication, etc.).In this process, the baby cannot be able to understand generalized knowledge laws through language, but can only carry out example and honorary processes such as "If (event a).)".In fact, the motor and cognitive skills of the child only exceed the current advanced AI models in most respects alone.
Of course you could tackle that "a baby could never learn through a test and a top -go -player or an expert in another field", as can be for people.As a result, algorithms teach.Instead of trusting very general laws, human "experts" are actually dependent on "tacit knowledge" - some shortcuts that our brains have learned from empirical information that appears at some point as inspiration, intuition or skills.
Figure 2 The Heuristic Learning of Alphazero.Lim: Simplified Go game tree in 5 x 5 table;
AI for real estate: we are still optimistic
We are in favor of accepting an approach that lacks explanations and trying to have made an effort about our understanding of reality with the help of AI, instead of limiting it with our obsession with interpretability.Ai who can understand complex interactions, well -covered statistical correlations and readyness.
The area of real estate is full of generalizations such as the cliché "location, location, location. But behind" location "are tens of thousands of factors that influence the real estate market, of demography of macro level, infrastructure, energy supply or climate change, on the most micro -Level of user experience, inside environmental quality and even what a person saw on social media a minute earlier, or the CO2 level in his/her brain.
AI algorithms themselves have of course the potential to do this.etc. at the same time created different AI algorithms that effectively use data to extract information.
Despite a bit of a setback, the Zillow stock price is still twice what it was when it started the Zillow Ibuying program, which is probably another factor that keeps us optimistic.
Reference:
Activation functions, and that are better in what task: Glorot, X., Bordes, A. and Bengio, Y. (2011b) .Dy Garry to the neural network.I JMLR W&CP: Course of the Fourteen International Conference on Artificial Intelligence and Statistics (Aistats 2011) .130, 297
How Alphago works: Pumperla, Max and Kevin Ferguson.Manning Publications Company, 2019 andhttps://nikcheerla.github.io/d...
Discussion about the Black-Box character of Deep Learning Models-Ssvar by Moenova from Toronto University:https://www.zhihu.com/question...
What can we learn from learning deephttps://mp.weixin.qq.com/s/vhG...
Wsj: What's wrong with Zillow?
The company had brought its future growth at its digital -flying company, but getting the algorithm turned out to be really difficult:https://www.wsj.com/articles/z...
Author (s)
- Alex Xudong Sun Ph.D.-Kandidat I Finansministeriet - Maastricht University
