[ 3 / biz / cgl / ck / diy / fa / ic / jp / lit / sci / vr / vt ] [ index / top / reports ] [ become a patron ] [ status ]
2023-11: Warosu is now out of extended maintenance.

/sci/ - Science & Math

View post   
File: 53 KB, 549x591, 1514917401116.png [View same] [iqdb] [saucenao] [google]
10155739 No.10155739 [Reply] [Original]

What are some of the biggest roadblocks and challenges in robotics and AI right now?

How can they be over come?

>> No.10155747

>>10155739
>robotics
Hard build hardware as precise human hands, control system and real world control.

>AI
AI moar layers, build powerful mathematical theory for Deep Learning and other methods.

>> No.10155756

>>10155747
On it

>> No.10155883

>>10155739
Model based reinforcement learning, unsupervised learning, real time inference and computer vision, agent planning and control, data simulation, data labeling, distribution shift, adversarial attacks, autonomous systems and vehicles, data privacy and ethics, etc.

>> No.10155888

>>10155747
If we need moar layers, why can't we just make networks that are 100 neurons wide and 1 million layers deep?

>> No.10155894

>>10155888
Because the number of parameters grows exponentially with depth.

>> No.10155905

>>10155894
Really? Why? Even if every neuron is connected to every other neuron, doesn't that grow only quadratically with depth?

>> No.10155945

>>10155905
if you keep the width for each layer the same, its quadratic. in practice, people usually double the number of hidden units in each layer to learn more complex features.